Invest3 publishers3 min readPublished Updated
Thomson Reuters keeps its win over ROSS in the first appellate test of AI-training fair use
Thomson Reuters kept its copyright win over ROSS Intelligence on appeal, the first US appellate ruling on fair use in AI training. ROSS had been refused a Westlaw licence as a rival, so the ruling lets owners of training data set a price or decline to sell.
The Investor · Invest desk
What happened
- Judge Stephanos Bibas revised his own 2023 decision in February 2025 and found ROSS infringed 2,243 Westlaw headnotes, rejecting its fair-use defence.
- Bibas said the court was considering "only non-generative AI", and ROSS had built a legal search engine, not a large language model.
- The US Copyright Office wants voluntary licensing markets to form before Congress considers compulsory measures, while conceding that licensing at training scale is costly and hard to manage.
Compiled by The InvestorSomething wrong?How this is made
Why it matters
- constraint A company refused a licence by a rival cannot count on a third party's derivative of that data to train a competing product, because the court treated that route as infringement.
- cost The licensing bill falls on whoever can afford it; Cryptopolitan and the OECD both expect that to favour companies already holding data, compute and cash.
- precedent Content owners that also sell products built on their data, as Thomson Reuters does with Westlaw, now have an appellate ruling behind a refusal to license a competitor.
- decision Congress's choice between voluntary and compulsory licensing now decides whether an owner can refuse a rival at all, since only a compulsory scheme would force a sale.
ROSS tried to pay before it copied. Thomson Reuters turned down the licence request because ROSS competed with it [3], so ROSS went to LegalEase and bought almost 25,000 Bulk Memos written from Westlaw headnotes [4]. The court judged that copying by what it was for. "Ross took the headnotes to make it easier to develop a competing legal research tool. So Ross's use is not transformative," Bibas wrote [7].
That sequence makes this an odd licensing precedent to price. Cryptopolitan argues that if AI firms have to pay more for training data, high-quality proprietary content becomes one more advantage for companies with big wallets [11], and the OECD says tight control of data, computing power and other critical inputs already favours the prominent players [12]. The money exists. Goldman Sachs expects global AI investment of about $1 trillion by 2026, with $581 billion of it in the US [13], or about 58% [1]. The US is also where these disputes are being settled one fair-use case at a time [14]. ROSS's problem was a seller that would not sell to a competitor, and a larger budget would not have changed that answer. The cost is a licence fee when the owner will deal, or rather, when the owner is also the competitor, the whole product the buyer meant to build. Cryptopolitan's account of the ruling, drawn from Reuters' September 29 report [1], does not include a damages figure or a licence fee, so that cost cannot yet be put in dollars.
The ruling could stay narrow in two ways, and Congress could override it in a third. The non-generative limit [8] leaves a model developer free to argue that a general-purpose model is a different use from a search engine built to replace its source. The Copyright Office weighs how material was obtained and how its use affects the value of the original [9]. ROSS's facts, derivative memos bought after a refusal and used for a rival product, score poorly on both [3] [4]. Congress could also impose the compulsory licensing that the Copyright Office wants voluntary markets to come before [10].
Europe has built a different system. Rightsholders there can reserve works from text and data mining under Article 4 of the 2019 DSM Directive [15], and in July 2026 the European Commission published a feasibility study for a registry that would help AI developers find those opt-outs through fingerprints, identifiers and metadata [16].
I think the practical effect in the US falls on products built to compete with their data source. Those will license the data or train without it. Two outcomes would prove that wrong: an appellate court finding generative training fair even where the output competes with the source, or a compulsory scheme that forces an owner like Thomson Reuters to sell to a rival at a set rate.
What to watch
- How the courts hearing the copyright fight over OpenAI and Microsoft treat the ROSS ruling's limit to non-generative AI.
- Whether the European Commission moves from its July 2026 feasibility study to building an EU registry of text-and-data-mining opt-outs.
- Any disclosed fee in a voluntary AI-training licence, the first number that would size what this ruling costs a buyer.